[Mesa-dev] [PATCH] i965/fs: Change the type of booleans to UD and emit correct immediates

[Jason Ekstrand]

On Tue, Oct 28, 2014 at 12:29 PM, Matt Turner <mattst88 at gmail.com> wrote:

> On Tue, Oct 28, 2014 at 12:10 PM, Jason Ekstrand <jason at jlekstrand.net>
> wrote:
> >
> > On Oct 28, 2014 11:57 AM, "Matt Turner" <mattst88 at gmail.com> wrote:
> >>
> >> On Thu, Oct 16, 2014 at 3:40 PM, Jason Ekstrand <jason at jlekstrand.net>
> >> wrote:
> >> > Before, we used the a signed d-word for booleans and the immedates we
> >> > emitted varried between signed and unsigned.  This commit changes the
> >> > type
> >> > to unsigned (I think that makes more sense) and makes immediates more
> >> > consistent.  This allows copy propagation to work better cleans up
> some
> >> > instructions.
> >> >
> >> > total instructions in shared programs: 5473519 -> 5465864 (-0.14%)
> >> > instructions in affected programs:     432849 -> 425194 (-1.77%)
> >> > GAINED:                                27
> >> > LOST:                                  0
> >>
> >> I assumed at first that this was on Haswell, but it couldn't be
> >> because Haswell doesn't use 0/1 for boolean. What platform was this?
> >
> > It doesn't matter what form of booleans the arch uses.  I believe it was
> on
> > HSW.
>
> It's unclear to me how you're coming to that conclusion. None of the
> hunks in brw_fs_visitor.cpp affect platforms where
> Const.UniformBooleanTrue != 1.
>

Not true.  The real problem we were hitting was when we emitted a value
from an ir_constant.  In this case, we emitted it as a copy from unsigned
to signed regardless of your platform.  That was what was causing us extra
instructions.  Most of the other changes are purely cosmetic because I
decided to unify on UD rather than D.


> I suppose the meaningful change in this patch for those platforms is
> the one in brw_shader.cpp.
>
> > However, as I have mentioned in private, I've had trouble running
> > shaderdb and believing the results.  They could be bogus but I don't
> think
> > so.
>
> Also unclear to me how you could not when I just tested it myself.
>

Re-running shader-db myself...

--Jason
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20141028/b493402a/attachment.html>